Qdisc layer. Fast enough for 10G wirespeed? Jesper Dangaard Brouer Hannes Frederic Sowa Daniel Borkmann Florian Westphal

Size: px

Start display at page:

Download "Qdisc layer. Fast enough for 10G wirespeed? Jesper Dangaard Brouer Hannes Frederic Sowa Daniel Borkmann Florian Westphal"

Joseph Dennis Hoover
6 years ago
Views:

1 Qdisc layer Fast enough for 10G wirespeed? Jesper Dangaard Brouer Hannes Frederic Sowa Daniel Borkmann Florian Westphal Network-Services-Team, Red Hat inc. 1/17 Netfilter Workshop, July 2014

2 Overview Analysing qdisc code path Estimating and measuring overhead Batching Results with bulk dequeue Lockless FIFO implementation Results and comparisons 2/17

3 Basic question? Basic problem/question: Is qdisc code path fast enough for 10G wirespeed? 10Gbit/s wirespeed smallest frame: Mpps -> Time Budget: 67.2 ns per packet 3/17

4 Locking: in qdisc code path Optimal "fast-path" where the qdisc is empty The following locking occurs: In dev_xmit_skb(): (1) spin_lock(root_lock) (detect qdisc empty, allow direct xmit) calling sch_direct_xmit() In sch_direct_xmit(): (2) spin_unlock(root_lock) In sch_direct_xmit(): (3) HARD_TX_LOCK(dev, txq, smp_processor_id()); In sch_direct_xmit(): (4) HARD_TX_UNLOCK(dev, txq); In sch_direct_xmit(): (5) spin_lock(root_lock); Return to dev_xmit_skb(): (6) spin_unlock(root_lock); LOCK cost on this arch: approx 8 ns 8 ns * 6 LOCK-ops = 48 ns pure lock overhead 4/17

5 Test: Remove qdisc from device Remove qdisc from device, like software/virtual devices qdisc path skipped, if (qdisc->enqueue == NULL) Run time hack (by Eric Dumazet) ifconfig $DEV txqueuelen 0 tc qdisc add dev $DEV root sfq tc qdisc del dev $DEV root Gotcha's Still takes the TXQ lock HARD_TX_{LOCK,UNLOCK} Thus, "only" avoiding 4 LOCK calls on fast-path Don't respect TXQ/device "full" state e.g. ignoring netif_xmit_frozen_or_stopped(txq) 5/17

6 Result: Remove qdisc from device Measurement done on CPU E5-2695(ES) with trafgen, packet config: udp_example01 Using qdisc NULL hack (Clean kernel rc6) pps = ns -- qdisc-path pps = ns -- qdisc NULL hack pps = ns -- improvement Theory: 4 locks saved = 32ns remaining 2 locks = 16ns Derived remaining cost of qdisc path = 12.5 ns Locking: Significant part of qdisc cost 6/17

7 Measured overhead vs. wirespeed Qdisc code path measured overhead: 60 ns Problematic for wirespeed 67.2 ns budget How can we reduce this overhead? Batching to amortize locking cost Lockless FIFO implementation 7/17

8 Difficult testing batching Must accumulate queue in qdisc Before batching of packets can happen Missing fast enough packet generator For single CPU overloading test Because default qdisc tied to HW TXQ queue Options External pktgen forwarding traffic Modify pktgen to send to qdisc layer Simply install single queue pfifo qdisc 8/17

9 Test setup Choose to simply use pfifo qdisc Mostly as quick solution Using trafgen, on multiple CPUs (option cpus) Remember use --qdisc-path option Trafgen packet config:udp_example01.trafgen fgen/udp_example01.trafgen Measurements done on CPU E5-2695(ES) 9/17

10 Results: pfifo baseline Baseline results, clean kernel rc6 Cpu 1: 1,570,083 pps baseline cpus /17

11 Implement: Bulk qdisc dequeue By bulk qdisc dequeue Amortize root lock, unlock+lock per N packets Actually helps enqueue'ers get root lock Reduce time dequeue blocks root lock for enqueue Still TXQ lock for every packet Due to SKB's have assigned a TXQ on enqueue 11/17

12 Results: Bulk qdisc dequeue Bulk dequeue results Cpu 1: 1,565,117 pps baseline bulk-dequeue cpus /17

13 Implement: Lockless FIFO qdisc Needs John Fastabend qdisc RCU work My hack just ignores stats Based on: Multi-Producer-Multi-Consumer ring queue Requiring LOCK'ed cmpxchg Bulk dequeue amortize cost of cmpxchg Enqueue cmpxchg on every packet Higher number of CPUs Hotspot/sync-point on cmpxchg retry loop 13/17

14 Results: Lockless FIFO qdisc Lockless FIFO with bulk dequeue Cpu 1: 1,726,467 pps baseline bulk-dequeue lockless cpus /17

15 Comparing default use-case: Single CPU Lockless compared to qdisc NULL hack pps = ns -- qdisc-path (clean kernel) pps = ns -- qdisc NULL hack pps = ns -- Lockless qdisc Difference against NULL qdisc: very close pps => ns Improvement against clean kernel: pps => ns Better than expected Likely due to ignoring stats 15/17

16 Lockless: Forcing through MPMC Want to measure clean overhead of MPMC queue Test: single CPU forced through MPMC By removing TCQ_F_CAN_BYPASS flag Expect added cost of two cmpxchg Results lockless qdisc w/o TCQ_F_CAN_BYPASS pps = ns -- with BYPASS pps = ns -- no BYPASS pps => ns -- Difference to bypass Were expecting to see 2x LOCK cost But only see approx 1x LOCK cost 16/17

17 The End Open discussion Where should we start? Need John Fastabend's qdisc RCU work Should we just bypass qdisc layer? We/Red Hat Will allocate resources/persons to project Open to collaborate 17/17

18 Extra Extra slides 18/17

19 Using TSQ TCP Small Queue (TSQ) Use queue build up in TSQ To send a bulk xmit To take advantage of HW TXQ tail ptr update Should we allow/use Qdisc bulk enqueue Detecting qdisc is empty allowing direct_xmit_bulk? 19/17

TX bulking and qdisc layer

Session: Linux packet processing performance improvements TX bulking and qdisc layer Jesper Dangaard Brouer (Red Hat) John Fastabend (Intel) John Ronciak (Intel) 1/16 Linux Plumbers Conference 16 th Oct